基于残缺图像样本的生成对抗网络图像修复方法

针对大面积图像修复缺失严重时,需要完整且高质量训练样本的问题,提出了一种将残缺或含噪图像样本作为训练集的双生成器深度卷积生成对抗网络（DGDCGAN）模型。构建两个生成器和一个鉴别器以解决单一生成器收敛慢的问题,用残缺图像样本作为训练集,通过交叉计算、搜索损失区域类似的图像信息作为训练生成模型的样本,收敛速度更快。鉴别器损失函数改进为输出的Wasserstein距离,使用自适应估计算法优化生成器损失函数和鉴别器损失函数的模型参数,最小化两两图像之间的总距离差,使用鉴别模型和修复图像总距离变化均方差最小化两个指标优化修复结果。在4个公开数据集上进行主客观实验,结果表明:所提方法能使用残缺图像样本作为训练集,有效实现大面积失真图像的修复,且收敛速度和修复效果优于现有图像修复方法。 

Image inpainting method based on incomplete image samples in generative adversarial network

A model of Double Generator Deep Convolutional Generative Adversarial Network (DGDCGAN), which uses the incomplete or noisy sample image as the training set, is proposed, in order to solve the problem of serious distortion of large area image inpainting, complete and high-quality training samples are frequently required, which is hard to acquire. Furthermore, the convergence of single generator is slow. Therefore, two generators and a discriminator are constructed. The incomplete image training set is used to cross calculate and search the image information similar to the loss area as the sample of training generation model, which achieves faster convergence speed. The loss function of the discriminator is improved to be the Wasserstein distance of the output. The adaptive estimation algorithm is used to optimize the model parameters for generating network loss function and identifying network loss function. Finally, the distance difference between two sets of images is calculated, and the reconstructed image is optimized by discriminating model and minimizing mean square error of the total distance change of a group of repaired images. Experiments are performed on four public dataset, the subjective and objective experimental results show that the proposed method that uses incomplete samples as training data can restore large area of distortion in images with faster convergence speed and better performance compared with the existing methods in image inpainting.